LoRA Stack Entry (Quantized)

LoRAs and quantized models have a rough relationship. When you load a model in INT8, W4A4, or W8A8 and then slap a normal Load LoRA node on it, ComfyUI has to dequantize each layer, patch it in floating point, and requantize it on every step - a known slowdown that gets worse the more LoRAs you stack. This node is the input end of the fix this pack ships for that exact problem. It's deliberately boring, and that's the point.

What it actually is

LoRA Stack Entry (Quantized) declares one LoRA for a stack. It doesn't load a model, and it doesn't patch anything - the class description is honest about this: it "configures one LoRA for Apply LoRA Stack (Quantized) without loading or modifying a MODEL." All it produces is a little spec object that gets handed to its sibling node. Think of it as the "in" box: you get one entry per LoRA you want in the stack, and the patcher downstream does the heavy lifting.

The workflow looks like this:

Load Diffusion Model Quantized
  -> LoRA Stack Entry (Quantized) --\  (one entry per LoRA)
  -> LoRA Stack Entry (Quantized) --/
        -> Apply LoRA Stack (Quantized) -> sampler

The inputs that matter

There are only two, and you'll set both every time:

• path - a dropdown of everything in your ComfyUI/models/loras folder, shown as a path relative to that folder. Pick your .safetensors here.
• strength - default 1.0, range -10 to 10, step 0.01. The tooltip spells out the useful part: set it to 0 or bypass the node to disable that LoRA.

That last bit is the quiet killer feature. Because each entry is an independent node, you can flip a LoRA off by zeroing its strength or bypassing it without deleting the connection and without touching the patcher. Stack four LoRAs, A/B one of them, leave the rest alone - it's the rgthree-style quality-of-life move this ecosystem is built on.

The output

One output, awkwardly named lora, with the custom type QUANTIZATION_TOOLKIT_LORA. That socket only plugs into Apply LoRA Stack (Quantized). Don't try to feed it into a stock loader - it's a config object, not a tensor.

Install

This is half of a two-node pack, so you're installing the whole thing: ComfyUI Quantization Toolkit (the current name for this fork). Easiest route is ComfyUI Manager - search "Quantization Toolkit". Or clone it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/SparknightLLC/ComfyUI-QuantizationToolkit

Then restart ComfyUI. Heads-up on the naming: this pack has been renamed a couple of times (formerly ComfyUI-INT8-Toolkit, and the immutable registry ID is still the older "ComfyUI-INT8-Fast-Fork"), so the folder and node IDs won't always match the display name you searched. It's the same thing. Requirements are modest - ComfyUI 0.32.0 or newer, a compatible comfy-kitchen, an NVIDIA GPU with usable INT8 throughput, optional Triton for the alternate INT8 backend - and the pack itself pulls in zero Python dependencies.

Gotchas

The dropdown only lists files inside your loras folder, so drop LoRAs there first or they won't show. And remember the ordering rule from the pack docs: if you're quantizing an existing workflow on the fly with Enable Quantization on MODEL, the recommended path is to load your stock LoRAs before quantization and bake them in, not to route them through this stack. This stack is for when the model is already quantized, or you want per-LoRA bypass control without re-baking. Get that backwards and you'll wonder why your images changed.